LZ LoRA Loader (Model Only)

Most LoRA loaders apply a patch to both the diffusion model and the CLIP text encoder - that's the "model + clip" weight pair you see everywhere. But a lot of LoRAs are purely visual: a style, a character's look, a lighting recipe. Patching the CLIP on those is pointless at best and slightly damaging at worst. LZ LoRA Loader (Model Only) strips the CLIP half out of the equation: model in, patched model out, no text encoder involved.

How it works

• model in - the MODEL from any checkpoint or diffusion-model loader.
• lora_name - dropdown of your installed LoRAs (a "None" entry is included).
• strength_model - the patch strength, range −10 to 10.

Internally it's the model-only branch of ComfyUI's LoRA application - it reads the LoRA file with load_torch_file and applies only the model-state-dict patch. If lora_name is "None" or strength_model is 0, it short-circuits and passes the model through untouched.

Outputs are MODEL plus two strings: lora_name and lora_strength. Those strings exist so you can wire "what got applied" into the pack's logging chain - LZSaveImageAndLog and the pipe system record them, which keeps your metadata honest about which LoRA and strength actually produced an image.

When it matters

Three realistic cases: style/color LoRAs that carry no trigger vocabulary (there's nothing for CLIP to learn, so skip the clip patch), chaining a model-only LoRA after something that already handled CLIP, and debugging - if your stacked output looks off and you suspect a LoRA is corrupting prompt understanding, this node isolates that variable. The tradeoff is real: if a LoRA does ship a trigger word or concept you want in the text encoder, this node won't apply it. Check whether your LoRA's metadata mentions a trigger phrase before assuming "model only" is right.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/liz-ils/ComfyUI-LZNodes

restart → MyCustomNodes/Loaders. ComfyUI Manager: search ComfyUI-LZNodes. No extra pip packages.

The bottom line

It's a small, honest utility - the kind of node you install the pack for and only use when you know why you're using it. If you're not sure whether a given LoRA should be model-only, the safe default is the full stacker or a normal LoRA loader; this is for when you understand the difference and want the precision. Used right, it's one less way for a LoRA to quietly fight your prompt.